Professional Documents
Culture Documents
OPEN ACCESS
Perspective
The Merging of Biological
and Electronic Circuits
Jonathan J.Y. Teo1,2,4 and Rahul Sarpeshkar3,*
SUMMARY
Biological circuits and systems within even a single cell need to be represented by
large-scale feedback networks of nonlinear, stochastic, stiff, asynchronous, non-
modular coupled differential equations governing complex molecular interac-
tions. Thus, rational drug discovery and synthetic biological design is difficult.
We suggest that a four-pronged interdisciplinary approach merging biology
and electronics can help: (1) The mapping of biological circuits to electronic cir-
cuits via quantitatively exact schematics; (2) The use of existing electronic circuit
software for hierarchical modeling, design, and analysis with such schematics; (3)
The use of cytomorphic electronic hardware for rapid stochastic simulation of cir-
cuit schematics and associated parameter discovery to fit measured biological
data; (4) The use of bio-electronic reporting circuits rather than bio-optical circuits
for measurement. We suggest how these approaches can be combined to auto-
mate design, modeling, analysis, simulation, and quantitative fitting of measured
data from a synthetic biological operational amplifier circuit in living microbial
cells.
INTRODUCTION
Biological networks are notoriously difficult to analyze and interpret because they comprise tens of thou-
sands of biochemical pathways that are linked. Pathways that are studied in isolation seldom provide deep
insights into the behaviors of larger networks because they do not account for the complex interactions
between pathways. In medicine, such systems biology interactions are important for drug or drug cocktail
discovery based on pathway analysis. They are necessary for predicting drug or drug cocktail efficacy,
safety, side effects, mechanism and site of action, and changes in cocktail composition or dosage with
changing symptoms.
Molecular interactions manifest in several forms. For example, they can include resource limitations, which
couple biochemical reactions that share the same enzyme or molecule; loading or non-modular effects that 1Computational and Systems
Biology Program at the
change the behavior of pathways depending on their downstream and upstream connections (Cardinale
Massachusetts Institute of
and Arkin, 2012); cross-talk between pathways caused by synergistic or antagonistic effects of transcription Technology, Cambridge, MA
and translation regulation (Mukherji and Van Oudenaarden, 2009; Zeng et al., 2018) or by molecules having 02139, USA
2Thayer
multiple fan-out interactions with downstream targets; feedback effects from interactions that form a School of
Engineering, Dartmouth
closed loop between upstream and downstream molecules (Ferrell, 2002). To compound these difficulties,
College, Hanover, NH 03755,
biochemical reactions are generally non-linear and noisy () such that simplified biological models that as- USA
sume linear and/or deterministic interactions are not accurate. 3Departments of
Engineering, Microbiology &
Immunology, Physics, and
One solution adopted by synthetic biologists is to fit biological circuits into neat computational blocks Molecular and Systems
defined by logic gates. This digital framework has seen some successes in designing small biological cir- Biology, Dartmouth College,
cuits that use a relatively small number of logic gates. For example, Ausländer et al. built a 1-bit adder in a Hanover, NH 03755, USA
4Present address: Genome
mammalian system using over 20 logic gates distributed among 9 different cell types (Ausländer et al.,
Institute of Singapore,
2018). So far, it has proven difficult to scale these circuits to larger networks due to issues such as resource Agency for Science,
limitations including metabolic burden, loading and non-modularity, stochastics (noise), cross-talk, and Technology and Research,
feedback (Sarpeshkar, 2014; Teo et al., 2015). 138672 Singapore
*Correspondence:
rahul.sarpeshkar@
Bio-molecular reactions are analog and stochastic by nature, so a blanket solution that treats all reactions dartmouth.edu
like logic gates works optimally only in special cases such as memory storage and decision-making where https://doi.org/10.1016/j.isci.
deterministic digital behavior is appropriate. Additionally, an analog approach offers resource (molecular 2020.101688
copy number) and energy (ATP) advantages compared to a digital approach for many biological functions
that are low precision by nature. For example, we showed in a previous paper (Sarpeshkar, 2014) that the
number of molecules needed to perform 2-bit precise addition is 3 orders of magnitude lower with an
analog approach compared to a digital approach. The energy advantage of the same addition is almost
4 orders of magnitude lower when using an analog approach. By adopting an analog electronic circuit
framework, we can use a language that describes biological processes more naturally. We also gain an
appreciation for the power, speed, and precision trade-offs exhibited by nature when she makes complex
circuits and systems function under tight resource constraints of power (ATP/sec) for a given speed (1/(time
taken to do the computation)) and precision (signal-to-noise ratio or bits of precision) (Sarpeshkar, 2014).
All of the problems that we have described with respect to biological circuits and systems are also ex-
hibited in analog electronic circuits, especially in those that have resource limitations with respect to en-
ergy and power and part count. For example, analog circuit designers also need to build robust functional
devices using components that are often noisy at low electron copy number and that can be subject to sig-
nificant manufacturing variations. Even well-characterized components sometimes exhibit unintended be-
haviors when operating in the context of a larger circuit due to loading, cross-talk, and feedback. Analog
circuit designers often build systems that operate within a range of robustness and performance specifica-
tions while subject to constraints of power or physical space on an integrated circuit (chip) or board. For
example, medical implants that are both small and power efficient (Bradley, 2006; Sarpeshkar, 2010) use
ultra-low-power electronics to operate. Such electronic design is often made possible by a deep under-
standing of feedback systems that circuit designers have developed and refined over the past 100 years
(Sarpeshkar 2010, 2014) for linear and nonlinear circuits.
Given the many similarities between biological and electronic circuits, we posited in a previous review pa-
per that 17 synthetic biological circuits over 20+ years could be effectively studied through the lens of
analog circuits for obtaining insight into their operation at the DNA, RNA, protein, or small-molecule/
metabolite levels (Teo et al., 2015). Figures S1 and S2 and associated captions summarize the key ideas
in this paper for the reader’s convenience. Intuitively, flow variables can be mapped to current; state vari-
ables can be mapped to voltage; relationships between these variables, whether linear or nonlinear, map
to corresponding current-voltage relationships as in Figure 3. Readers interested in detailed mathematical
derivations of biochemical reaction networks can also find them in chapter 24 of (Sarpeshkar 2010) with
associated hardware work described in (Woo et al. 2015, 2018; Kim et al., 2018). Important relationships
between energy, information, and power in cells can be found in (Sarpeshkar 2014). Here, we suggest a
four-pronged interdisciplinary methodology that is capable of leveraging the full power of analog elec-
tronic circuit design for general modeling, design, analysis, rapid simulation, and electronic reporting
and measurement of biological circuits as shown in Figure 1. The Results subsections of our paper are orga-
nized as follows. In Section I, we discuss how to create electronic circuit schematics that are an exact rep-
resentation of mathematical equations that describe biological circuits and use the enzyme-substrate bind-
ing circuit as an example (Figures 2 and 3). In Section II, we show how existing electronic circuit software,
e.g. Cadence, which has been used for the hierarchical design of billions of electronic transistor devices on
modern complex integrated circuit chips for decades, can be leveraged for the design of biological circuits.
To show how our method can work in concrete detail and practice, we focus on a synthetic microbial bio-
logical operational amplifier circuit as shown in Figure 4. In Section III, we discuss how custom cytomorphic
integrated circuit chips can rapidly simulate many instantiations and parameters of biological circuits that
have been compiled into equivalent electronic circuits and thus help with automated parameter discovery
and automated data fitting of measured biological data. We show a concrete example w.r.t. to the syn-
thetic biological operational amplifier (Figures 5, 6, and 7). In Section IV, we illustrate how bio-electronic
rather than bio-optical reporting and measurement can further help with the merging of biology and elec-
tronics with respect to such automation (Figure 8). We conclude by discussing the potential impact of our
work in the future. Supplemental Sections I-III provide further details that correspond to Sections II and III in
the main body of the paper.
system complexity scales, they are difficult to interpret meaningfully. We have developed a mapping from
biology to electronics that combines the ease of interpretation of illustrations with the mathematical exact-
ness of ODEs. It can be applied to all circuits in molecular biology at the DNA, RNA, protein, and small-
molecule levels, which we have illustrated as a canonical circuit schematic in a previous review (Teo
et al., 2015) and in Figure S2. To demonstrate how the general mapping works, outlined in Figure 3, we
use the specific and ubiquitous example of an enzyme-substrate (ES) binding reaction shown in Figure 2.
ES binding is typically represented by a cartoon or a mass action equation as in Figure 2. The cartoon pro-
vides a quick visual description while the equations capture the underlying thermodynamics. The equations
in Figure 2 show that the rate of association of E and S depends on the product of free enzyme and free
substrate multiplied by a forward rate constant Kf, while the rate of dissociation depends on the concen-
tration of bound complex with a reverse rate constant Kr. We can represent such dynamics using a circuit
where a current generator, resistor, and capacitor are connected in parallel. Such a pictorial circuit repre-
sentation allows biological circuit designers to identify key characteristics such as feedback and noise
without performing laborious calculations: Current fluxes through the current generator and resistor repre-
sent the association and dissociation molecular fluxes, respectively. As current flows in and out of the
capacitor, its voltage changes accordingly. The voltage across the capacitor in this case corresponds to
the concentration of bound substrate complex. The circuit subtracts bound variables from total conserved
molecular variables to create free voltage variables via the voltage subtractors, creating two parallel ‘‘use-
and-lose-it’’ negative feedback loops that are ubiquitously present in all chemical reactions. These ‘‘free
voltage variables’’ are multiplied to generate a current that matches the forward association flux. The over-
all analog schematic is one that represents biological interactions succinctly while being faithful to the un-
derlying mathematical descriptions including noise, dynamics, robustness, and feedback (Sarpeshkar,
2010, 2014; Teo et al., 2015). Figure S1, adapted from (Teo et al., 2015), provides further details.
Figure S1 shows that even diffusion can be exactly modeled via distributed resistor-capacitor (RC) circuits.
Chapter 14 in (Sarpeshkar 2010) discusses how current-mode diffusor circuits can map such distributed RC
circuits to highly efficient current-mode equivalents suitable for integrated circuit implementation, e.g., on
the cytomorphic chips, discussed in Section III. In practice, as shown in Figure S5 and (Woo et al. 2015,
2018), cytomorphic chips and boards are assembled with field programmable gate arrays (FPGAs) for
digital programmability of network connectivity and analog parameters. Therefore, via simple attenuation
and delays in digital data packets to adjacent addresses, delays and diffusion can also be easily imple-
mented in the digital domain.
To uncover properties of a biological circuit, we can study its analog schematic equivalent using analytical
tools from analog circuit design. In the case of an ES binding circuit, the Poisson noise in ES flux is effec-
tively the sum of Poisson noise from forward and reverse current fluxes in the dependent current generator
and resistor, respectively, and determines the voltage or concentration noise on the capacitor (Sarpeshkar
2010, 2014). If the substrate negative feedback loop is effectively absent due to a large concentration of
substrate, the enzyme negative feedback loop with a small-signal loop gain of Kf STOT/Kr = S/KD determines
the proportional tracking of ES (output) with ETOT (input) as in a simple negative feedback buffer opera-
tional amplifier circuit in electronics, generating well-known Michaelis-Menten input-output relationships.
Figure 3 summarizes how mappings can be similarly created for 10 other common bio-molecular processes.
The papers (Sarpeshkar, 2014; Teo et al., 2015) provide further details on the same treatment for several
other well-known circuits from synthetic biology.
Section II: Using Electronic Circuit Software for Biological Circuit Design
Circuits, which convert equations to pictures, enable a big picture intuitive view of a whole system with all of
the important feedback loops and interactions visibly obvious in a schematic or map. It is possible to design
small circuits without such a schematic by using only mathematical equations. But, for design, analysis, and
simulation of circuits and systems with hundreds to millions of state variables, pictorial intuition, model-or-
der reduction, and hierarchy are essential. It is also important that we have the ability to rapidly switch the
complexity of a model in a simulation to see if it actually matters at the timescale and for input-output re-
lationships that are of relevant experimental biological importance in a practical application. For example,
a ‘‘level 1’’ model with two experimentally constrained parameters may be more useful in a practical
application even though the actual biology may involve significantly more parameters in an accurate ‘‘level
57’’ model with 1,000s of free parameters, many of which are unknown and are not experimentally con-
strained or relevant. Thus, certain subsystems can be well approximated with reduced-order input-output
circuits: Finally, analog schematics are important for the scaling of biological circuits from simple low-level
circuit motifs to circuits that can be connected and composed to create large-scale systems that are
composed in a hierarchical fashion: For example, many ES reactions in cascade can form a glycolysis circuit,
which itself can be a sub-circuit in a higher level glucose-to-ATP circuit. The glucose-to-ATP circuit can be a
metabolic driving circuit with an effective nonlinear driving source impedance that powers other small and
large subsystems within a cell. Many such subsystems can form the system of the cell.
Hierarchical and complex circuit design can be ported from electronic to biological circuit design with ex-
isting software. Over about 50 + years, analog circuit designers have developed computer-aided design
software such as Cadence and circuit simulation tools such as SPICE to visualize circuits and to simulate
them with ease (see Figure S3). SPICE also comes packaged with over 25 analysis tools, e.g. to perform
small-signal .AC frequency analysis; to study transient .tran time-domain behavior; to study steady-state
.DC input-output behavior; to see the effects of noise via .PNOISE; or to use monte-carlo simulation to
study parameter sensitivity. Such tools can be leveraged rather than re-inventing the wheel from scratch.
Schematic circuits at one level of a hierarchy can be connected to create a higher-level input-output circuit
at the next level of hierarchy, which is represented abstractly by a circuit ‘‘symbol’’. Such symbols can then
be combined with each other to create higher level circuits. The process can be recursively repeated to
create systems of massive complexity. The microprocessor and analog cell phone chips that we take for
granted today, which often have billions of interconnected transistor devices, were designed in such a
Figure 4. Continued
the resulting complex inhibits either LuxI or AiiA expression in the second stage. The concentration of aTc regulates
dCas9 expression, which enables control of the total strength of the inhibition. The last amplification stage is a fast
stage that uses LuxI and AiiA to synthesize or degrade AHL, respectively. AHL binds to LuxR at the inverting input to
close the feedback loop.
(C) To represent the production of bio-molecules in its stages of amplification, the schematic uses a circuit motif similar to
the enzyme-substrate binding motif in Figure 2.
fashion. Here, we shall only provide an example of a relatively simple synthetic biological operational
amplifier circuit to illustrate how electronic circuit software can be concretely used for creating biological
circuit schematics. Nevertheless, it will illustrate all of the concepts. Subsequently, we shall focus on discus-
sing how to compile, simulate, discover parameters, and fit data in an automated fashion.
Synthetic biologists have designed and ported many basic electronic circuit motifs such as logic gates
(Bonnet et al., 2013; Maung and Smolke, 2008; Nielsen et al., 2016), latches (Gardner et al., 2000; Lee
et al., 2016), load driver circuits (Mishra et al., 2014), oscillators (Chen et al., 2015; Elowitz and Leibier,
2000; Prindle et al., 2014), and log-linear circuits (Daniel et al., 2013) to biology. In contrast, the operational
amplifier (OpAmp), which is a popular device used in analog circuits, has only just started gaining traction
(Klavins, 2014). An OpAmp is often considered a fundamental building block for designing large regulatory
loops in analog circuits due to its many attributes for modulating feedback robustly. It implements a simple
circuit building block wherein the voltage difference of its two control inputs is amplified with a large gain
and reported at its output terminal (see Figure 4 and Appendices from (Zeng et al., 2018)). When connected
to companion components in a negative feedback loop, an OpAmp enables robust and precise analog
computations such as weighted summation, inverting and non-inverting amplification, and integration. It
performs these calculations rigorously despite being susceptible to manufacturing variations. An OpAmp’s
robustness to parameter variation is essential for making analog electronics that are scalable, versatile, and
context independent from an input-output point of view. Therefore, by building a biological OpAmp (Bio-
OpAmp) (Zeng et al., 2018), we can leverage its properties to regulate feedback in a variety of important
biological processes such as homeostasis (Berridge et al., 2003; Nemazanyy et al., 2015; Wong et al.,
2012), metabolic regulation (Pan et al., 2009; Sturm et al., 2010), and immune response (Fukao et al.,
2002; Liu et al., 2015).
The Bio-OpAmp design shown in Figures 4A–4C is a concrete instantiation of the framework of Figure 1. It
is largely inspired by a classical three-stage operational amplifier (Zeng et al., 2018). An electronic OpAmp
typically has three gain stages: an initial high-bandwidth input stage, an intermediate amplification stage,
and a high-bandwidth output stage. The Bio-OpAmp mimics this architecture using a combination of tran-
scription and translation regulatory mechanisms (See Figure 4B). The first stage is a fast differential input
stage that senses the difference between arabinose (non-inverting signal) and acyl homoserine lactone
(AHL) (inverting signal) and generates corresponding guide RNAs (sgRNAs). The sgRNAs compete for
binding with dCas9, such that when sgRNA1 binds to dCas9, it removes a potential binding site for sgRNA2
and vice versa, thereby amplifying the input signal via a ‘‘push-and-pull’’ mechanism. The second stage is a
slow differential stage that uses bound dCas9-sgRNA complexes to downregulate AiiA and LuxI protein
expression levels. Since each mRNA transcript of AiiA and LuxI is translated to multiple copies of their
respective proteins, changes to sgRNA levels are amplified via this downstream increase in copy number.
Translation generally happens on a slower timescale than transcription, so the second stage is slower to
respond to changes. The third stage is a fast differential stage where AiiA and LuxI enzymes degrade or
produce AHL, respectively. Since each enzyme can potentially catalyze multiple copies of AHL, the third
stage creates yet another level of amplification. Additionally, since both enzymes catalyze production/
degradation of the small molecule, they effectively combine the two inputs into a single-ended output.
When the three gain stages are combined, they produce a large open-loop gain that is difficult to achieve
using only a single gain stage in isolation due to biological limitations.
Further details of the Bio-OpAmp w.r.t use of electronic circuit software for biological circuit design
including robustness, sensitivity to parameters, buffering capability, oscillation, and tracking performance
are described in Supplemental Section I of this paper. Figures S10–S14 and Figure 7 show several details of
its performance, comparing circuit software and circuit hardware instantiation on cytomorphic chips. We
shall now focus on an important part of the framework of Figure 1, namely the compilation of circuit sche-
matics to cytomorphic chips.
Figure 5. Computational Pipeline to Program Cytomorphic Chips and Systems for Automating Parameter
Discovery and for Learning
Using compilers that have been developed or are currently in development (Medley et al., 2020), we translate biochemical
reactions in SBML or in SPICE netlists to programming bits for the cytomorphic chips that have hardware instantiations of
chemical reaction circuits. To aid in such programming, MATLAB and Simulink interfaces interact with field
programmable gate arrays (FPGAs) on circuit boards composed of many such cytomorphic chips. The programmable bits
configure DACs, switches, and bio-molecular network connectivity on the cytomorphic chips.
Section III: Compiling Circuit Schematics to Cytomorphic Chips for Fast Simulation,
Parameter Discovery, and Data Fitting
While existing circuit software tools like Cadence are useful for the design of large-scale systems and the
simulation of relatively small circuits like the Bio-OpAmp, the simulation of large-scale biological systems
can be slow on general-purpose digital computers. In particular, stochastic stimulations of relatively low-
copy-number molecular variables are slow even for systems with as few as 10 variables (Gibson and Bruck,
2000; Li and Petzold, 2010; Ramaswamy et al., 2009). Such simulations require Poisson process emulation
via the Gillespie stochastic simulation algorithm (GSSA), which is extremely computationally intensive on
digital computers, regardless of the software package used (MATLAB, Cadence, COPASI, ..) (Kim et
al., 2018; Woo et al., 2015). They are not parallelizable due to the need to faithfully emulate asynchronous
Poisson processes and create pseudo-random numbers at every time step on synchronous digital com-
puters with fixed time steps and clocks. Stiff differential equations with diffusion and compartmentalization
exacerbate such problems.
Fortunately, deterministic and stochastic simulations of large-scale bio-molecular dynamical systems can
be done extremely quickly on digitally programmable highly parallel asynchronous analog cytomorphic
systems (Sarpeshkar 2010; Kim et al., 2018; Woo et al., 2015, 2018). These systems have a flat simulation
time even as reaction network size or the number of species scales. Some of these systems amplify natural
thermal noise to achieve highly stochastic simulations or use inherent analog noise in current fluxes for
moderate levels of noise (Kim et al., 2018). Their readouts and programmability are compatible with tradi-
tional digital computers such that the user can be agnostic to the details of internal chip operation,
viewing such simulations as simply originating from a fast custom Bio co-processor. While there is an
argument to be made for keeping model complexity low such that data can be interpreted more mean-
ingfully, biology as a whole is moving toward understanding bigger and more complex systems. Knowl-
edge of such systems is rapidly progressing because of advances in next-generation sequencing technol-
ogy and because of the advent of big data and machine learning. Thus, it is advantageous to compile
circuits to cytomorphic chips for fast simulation and then leverage cytomorphic chip simulations to train
large biological networks that fit biological data. In this paper, we shall only briefly summarize how these
cytomorphic chips operate. Readers interested in further details should consult past work over more than
a decade (Sarpeshkar et al., 2005a; Mandal and Sarpeshkar, 2009a, 2009b; Sarpeshkar, 2010; Kim et al.,
2018; Woo et al., 2015, 2018).
As described extensively in the latter papers, the same equations of Boltzmann exponential thermodynamics
govern the stochastics of molecular reaction dynamics as well as current flow in subthreshold electronic tran-
sistor circuits. Thus, a mapping of the mathematical differential equations that govern interactions between
molecules with log (molecular concentration) mapped to voltage and molecular flux mapped to current,
respectively, lead to an efficient and exact translation of nonlinear biological circuits and systems to equivalent
nonlinear electronic circuits and systems that simulate them. Such physical emulation to do fast simulation is
analogous to the idea of using GHz inductor-capacitor-resistor (LCR) electronic circuits to simulate slow
spring-mass-damping mechanical circuits. In the papers (Kim et al., 2018; Woo et al., 2015, 2018), the authors
describe the architecture, specifications, and performance of the cytomorphic chips in detail. The authors
demonstrate the ability of the chips to be digitally configured to simulate any biochemical reaction network
using programmable circuit building blocks. For example, we can simulate synthesis, degradation, associa-
tion, dissociation, dimerization, substitution, cascade, fan-in, fan-out, and loop networks by appropriately
configuring and connecting analog circuit blocks. Figure S5 shows the key circuit building block, and Figures
S6B and S6C provide specific examples of a branching network important for competitive drug binding and ES
reactions, respectively. Other examples that have been described include the dynamics and stochastics of a
repressilator, a p53-MDM2 cancer pathway, and glycolytic oscillations. Even for modest networks with only
about 80 stochastic reactions, cytomorphic simulations exhibit a significant 700X speedup over digital COPASI
simulations or 30,000x speedup over digital MATLAB simulations, while yielding identical results (Woo et al.,
2015, 2018). For large-scale systems, such speedups could be greater than a million fold (Woo, 2016).
Building on such prior work, we were motivated to compile and simulate the Bio-OpAmp with cytomorphic
chips. To do so, we developed a mapping between analog circuit representations useful in this particular
Bio-OpAmp case and circuits on cytomorphic chips. The Bio-OpAmp uses three primary motifs: a transcrip-
tion-translation motif, a fan-out motif for substrate competition, and a dependent production and degra-
dation motif for AHL regulation. Figures S7–S9 and Supplemental Sections I and II discuss these mappings
in greater detail. The circuit in Figure 4C is compiled to its cytomorphic equivalent in Figure 6 using these
mappings. Specifically, we use the translation-transcription cytomorphic block from Figure S7 in OpAmp
gain stages 1 and 2; we use the conductance divider cytomorphic block that models sgRNA and dCas9
binding from Figure S8 in OpAmp gain stage 1; we use the AHL production-degradation cytomorphic
block from Figure S9 in OpAmp gain stage 3.
The pipeline shown in Figure 5 enables us to program the chip and run multiple simulations of the Bio-
OpAmp in parallel. The pipeline can adapted to read SBML files using a compiler that was developed
recently by Medley et al. (2020). It is worth noting that digital calibration can drastically improve the
precision and variability of analog circuits (Sarpeshkar, 2010). Calibrating analog-to-digital converters
(ADCs), digital-to-analog converters (DACs), and other analog circuits to yield high-precision, low vari-
ability analog systems has been proven in cochlear implants for deaf patients (Sarpeshkar et al, 2005a,
2005b) and in other applications (Sarpeshkar, 2010). Such calibration would be needed in actual commer-
cial or large-scale systems. However, they do not alter the conclusions of our smaller-scale proof-of-
concept demonstrations here such that we shall not focus on them.
Figures 7A and 7B show that Cadence software simulations and cytomorphic chip simulations of the Bio-
OpAmp are in good agreement, illustrating how the general framework of Figure 1 can be concretely
instantiated in practice. Using the pipeline of Figure 5, we programmed multiple chips to simulate the Bio-
OpAmp in parallel as shown in Figure S10. Chip-to-chip variations in such simulations can be calibrated for
digitally as in a past cochlear implant for the deaf, which worked on a deaf subject on the first try (Sarpesh-
kar, 2010; Sarpeshkar et al., 2005b).
We simulated the Bio-OpAmp in electronic circuit software and on cytomorphic chips with varying concen-
trations of AHL and arabinose. These simulations replicate the biological experiments in (Zeng et al., 2018)
that characterize the open- and closed-loop behavior of the Bio-OpAmp: As shown in Figure S11, the
steady-state AHL concentration (reported as RFP in the experiments) tracks input arabinose closely while
rejecting the added disturbance input AHL in closed-loop configurations. In an open-loop configuration,
output AHL exhibits non-linear input-output characteristics with input arabinose that are reminiscent of the
saturation function used to describe cooperative binding. The steady-state outputs from biological data,
software simulation, and cytomorphic chip implementation are in good agreement, suggesting that the
schematic and cytomorphic circuit models of Figures 4 and 6 are indeed accurate.
We ran high-throughput cytomorphic simulations to perform sensitivity analysis and parameter discovery to
optimize the Bio-OpAmp. For example, the open-loop gain is maximized when the dissociation constants
of sgRNA1 and sgRNA2 are similar, as shown in Figure S12. This finding illustrates the importance of symmetry
in all feedback systems and high-performance circuits. The closed-loop gain in Figure S13 is also in accord with
predictions of such gain from small-signal analysis corresponding to Figure S4 or (Zeng et al., 2018). Figure S14
shows that the steady-state tracking error is also in accord with the overall gain of the OpAmp as predicted
from feedback system theory. Chapter 5 of (Teo, 2019) provides a more detailed discussion including addi-
tional findings based on parameter discovery and machine learning. A short conference paper (Teo et al.,
2019) presents a highly abbreviated version of some portions of this journal paper.
In biological systems, due to low molecular copy numbers and/or high burst noise factor, the dominant
measured noise is of thermal origin, has a flat or ‘‘white’’ power spectral density, and obeys the Poisson
statistics of uncorrelated current flow at the fundamental chemical reaction level (Chapter 7 on noise in Sar-
peshkar 2010; Sarpeshkar 2014). Hence, as in the GSSA, we have mostly focused on only discussing the
emulation of such Poisson noise, which is fundamental to all thermodynamic processes. However, at the
system level, due to correlations, explicit feedback, filtering, self-organization, scale invariance, and ‘‘de-
fects’’ in biological membranes and other biological structures, one can also see ‘‘pink noise’’ with a sloping
power spectrum at frequencies below a certain relatively low corner frequency. Such noise is observed in
electronic transistors as well due to ‘‘traps’’ or ‘‘mobility fluctuations’’ in current flow where it is often called
‘‘1/f noise’’ or ‘‘power-law noise’’. From quantitative models of such power-law noise in devices and circuits
as discussed in chapters 7, 8, 12, 13, 14, and 24 in (Sarpeshkar 2010), we can certainly also emulate such
biological noise both explicitly, e.g., as the noise in the membrane, and implicitly, e.g., if it emerges in cir-
cuits due to time-scale invariance in certain systems, e.g. heartbeats and traps in transistors. Such emula-
tions can be done in both circuit software and cytomorphic chip hardware if desired. In fact, most electronic
circuit software has such 1/f noise models by default to ensure that electronic circuits perform as desired.
As discussed in chapter 19 in (Sarpeshkar 2010), emergent properties such as ‘‘stochastic resonance’’ wherein
the detectability of a signal is improved if an optimal amount of noise helps it cross a fixed threshold can
emerge in both electronics and biology and can also be easily represented and modeled. Such effects may
be present in p53 in cancer networks, among the networks that we have been able to successfully model. In
fact, we can even quantitatively model fundamental and system noise in any circuit including the Bio-OpAmp.
However, as in several synthetic biological circuits to date, the Bio-OpAmp is implemented with relatively high
copy numbers of molecules with many cells in solution. Thus, Bio-OpAmp open-loop gains were nearly always
stably measured to be in the 50–100 range, both in experimental biological measurements and in quantita-
tively accurate models that fit such biological data (Zeng et al., 2018). While current optical reporting and
measuring systems in biology, which are not prohibitively expensive, cannot easily and non-destructively mea-
sure signal and noise in a single cell easily, the more sensitive bio-electronic systems, which we describe in Sec-
tion IV, may be capable of such measurements in the future. Hence, we shall discuss them now.
to accelerate and simplify the circuit prototyping process, the testing and validation process is still based on
making optical measurements of a fluorescent or luminescent reporter molecule produced by the circuit.
Many modern methods such as fluorescent microscopy, flow cytometry, and short-read sequencing make
use of fluorescent proteins for their core functions (Bentley et al., 2008; Lichtman and Conchello, 2005; Lyons
and Parish, 1994). Optical biosensors are used in many canonical circuits in synthetic biology (Elowitz and Leib-
ier, 2000; Stricker et al., 2008), and they continue to be the default mode of reporting in most biological circuits.
However, despite their popularity, optical biosensors present several caveats; fluorescent proteins cause cyto-
toxicity when produced in large concentration (Shen et al., 2017); fluorophores undergo photobleaching when
exposed to long periods of excitation (Greenbaum et al., 2000) reducing dynamic range and limiting the num-
ber of consecutive measurements; long photoexcitation causes phototoxicity by generating reactive oxygen
species that damage the cell (Ganini et al., 2017). Due to these limitations, fluorescent biosensors are unsuit-
able for reporting certain types of behavior such as taking continuous readings in time-lapse microscopy,
which may be necessary for a DBTL paradigm.
To create biosensors for a wider range of circuits, we need biosensors that can take real-time measure-
ments, have a wide dynamic range and good sensitivity, and do not cause significant cell death. As an
example, we developed a microbial fuel cell (MFC) that uses an electricigenic bacterium to generate cur-
rents based on circuit activity (Zeng et al., 2019). The fuel cell consists of a co-culture of E. coli MG1655 and
S. oneidensis MR-1 that resides on the anode. A co-culture system was chosen because it exhibits greater
tolerance to metabolic burden and cytotoxicity and experiences less cross-talk compared to monoculture
systems (Lisa et al., 2014; Stephens et al., 2019). E. coli converts lactose in its environment to lactate with the
aid of an intermediate reporting enzyme LacZa while S. oneidensis metabolizes the resulting lactate and
transfers the electrons to the anode using either direct electron transfer or metabolite (e.g. flavin)-medi-
ated electron transfer (Busalmen et al., 2008; Okamoto et al., 2013, 2014). When we measure current in
the MFC, we are effectively measuring the concentration of LacZa in E. coli. To use the MFC as an electrical
biosensor, we simply replace fluorescent protein genes on the plasmid with LacZa and measure electrical
current instead of fluorescence intensity.
To generate an irreversible digital output, the comparator of Figure 8 uses LacI to repress TetR and
vice versa, thereby creating a switch-like behavior. Arabinose activates LacI expression via LuxR
while an inducer with action similar to lactose (IPTG) represses LacI activity. The relative strength of
the two antagonistic effects decides the final reporter output. When using a fluorescent reporter (RFP),
changes to IPTG concentration did not significantly affect the threshold concentration of arabinose
required to switch the comparator on. In contrast, when using a LacZa reporter, a higher IPTG concen-
tration increases the threshold concentration of arabinose needed to flip the comparator (See (Zeng
et al., 2019)). The wider dynamic range can be explained by LacZa catalyzing lactose to lactate, which
serves as a mode of signal amplification, improving both the measurement as well as providing a conve-
nient electronic output.
CONCLUDING REMARKS
We have suggested how electronic circuit design and measurement can serve to automate design,
modeling, analysis, simulation, and quantitative fitting of measured data as shown in the framework of Fig-
ure 1. We have also shown how such a framework can be concretely instantiated in a synthetic biological
operational amplifier circuit in living microbial cells. While this work is in a proof-of-concept and founda-
tional stage, in the future, large-scale biological circuits and systems for drug cocktail discovery could
be designed and simulated quickly for a systems biology and medical application. Alternatively, precise
and robust synthetic circuits for medicine, e.g, Bio-OpAmp-like homeostasis circuits for drug dosage con-
trol can benefit from the robust and predictive power of design automation, enabling quick applications in
multiple contexts rather than via tedious time-consuming experimentation. In both cases, the merging and
unification of biology and electronics via unifying and hierarchical circuit motifs, accurate modeling and
compilation, fast simulation for parameter discovery and learning, and bio-electronic measurement may
help scale the design and understanding of biological systems. Thus, the current state, which is to do rela-
tively simple or empirical design, could grow toward more complex and more rational design, which is
highly important if medicine and bioengineering are to scale.
METHODS
All methods can be found in the accompanying Transparent Methods supplemental file.
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.isci.2020.101688.
ACKNOWLEDGMENTS
We acknowledge useful discussions with Ji Zeng, Jaewook Kim, and Sung Sik Woo. This work was sup-
ported by the AFOSR under grant number FA9550-18-1-0467 and by the NIH under grant number R01
GM 123032-01 to R.S. J.T. was supported in part by funds from the A*STAR and by funds from Dartmouth
College.
AUTHOR CONTRIBUTIONS
J.T. and R.S. designed the model and the computational framework. J.T. carried out the implementation
and analyzed the data. J.T. and R.S. contributed to the writing of the manuscript. R.S. was in charge of over-
all direction and planning of the project.
REFERENCES
Ausländer, D., Ausländer, S., Pierrat, X., devices. IEEE 2006 Biomed. Circuits Syst. Conf. Daniel, R., Rubens, J.R., Sarpeshkar, R., and Lu,
Hellmann, L., Rachid, L., and Fussenegger, M. Healthc. Technol. Biocas 2006, 158–161. T.K. (2013). Synthetic analog computation in
(2018). Programmable full-adder computations in living cells. Nature 497, 619–623.
communicating three-dimensional cell cultures. Busalmen, J.P., Esteve-Núñez, A., Berná, A., and
Nat. Methods 15, 57–60. Feliu, J.M. (2008). C-type cytochromes wire Elowitz, M.B., and Leibier, S. (2000). A synthetic
electricity-producing bacteria to electrodes. oscillatory network of transcriptional regulators.
Bentley, D.R., Balasubramanian, S., Swerdlow, Angew. Chem. Int. Ed. 47, 4874–4877. Nature 403, 335–338.
H.P., Smith, G.P., Milton, J., Brown, C.G., Hall,
K.P., Evers, D.J., Barnes, C.L., Bignell, H.R., et al.
(2008). Accurate whole human genome Carbonell, P., Jervis, A.J., Robinson, C.J., Yan, C., Ferrell, J.E. (2002). Self-perpetuating states in
Dunstan, M., Swainston, N., Vinaixa, M., signal transduction: positive feedback, double-
sequencing using reversible terminator
chemistry. Nature 456, 53–59. Hollywood, K.A., Currin, A., Rattray, N.J.W., et al. negative feedback and bistability. Curr. Opin.
(2018). An automated Design-Build-Test-Learn Cell Biol. 14, 140–148.
pipeline for enhanced microbial production of
Berridge, M.J., Bootman, M.D., and Roderick,
fine chemicals. Commun. Biol. 1, 1–10. Fukao, T., Tanabe, M., Terauchi, Y., Ota, T.,
H.L. (2003). Calcium signalling: dynamics,
homeostasis and remodelling. Nat. Rev. Mol. Cell Matsuda, S., Asano, T., Kadowaki, T., Takeuchi, T.,
Biol. 4, 517–529. Cardinale, S., and Arkin, A.P. (2012). and Koyasu, S. (2002). P13K-mediated negative
Contextualizing context for synthetic biology - feedback regulation of IL-12 production in DCs.
Bonnet, J., Yin, P., Ortiz, M.E., Subsoontorn, P., identifying causes of failure of synthetic biological Nat. Immunol. 3, 875–881.
and Endy, D. (2013). Amplifying genetic logic systems. Biotechnol. J. 7, 856–866.
gates. Science 340, 599–603. Ganini, D., Leinisch, F., Kumar, A., Jiang, J.J.,
Chen, Y., Kim, J.K., Hirning, A.J., Josic, K., and Tokar, E.J., Malone, C.C., Petrovich, R.M., and
Bradley, P.D. (2006). An ultra low power, high Bennett, M.R. (2015). Emergent genetic Mason, R.P. (2017). Fluorescent proteins such as
performance Medical Implant Communication oscillations in a synthetic microbial consortium. eGFP lead to catalytic oxidative stress in cells.
System (MICS) transceiver for implantable Science 349, 986–989. Redox Biol. 12, 462–468.
Gardner, T.S., Cantor, C.R., and Collins, J.J. compiler for biological networks on silicon chips. Sarpeshkar, R., Baker, M.W., Salthouse, C.D., J.
(2000). Construction of a genetic toggle switch in PLoS Comput. Biol. 16, e1008063. Sit, Turicchia, L., and Zhak, S.M. (2005b). An
Escherichia coli. Nature 403, 339–342. analog bionic ear processor with zero-crossing
Mishra, D., Rivera, P.M., Lin, A., Del Vecchio, D., detection. In International Solid-State Circuits
Gibson, M.A., and Bruck, J. (2000). Efficient exact and Weiss, R. (2014). A load driver device for Conference, (IEEE International Digest of
stochastic simulation of chemical systems with engineering modularity in biological networks. Technical Papers. Solid-State Circuits
many species and many channels. J. Phys. Chem. Nat. Biotechnol. 32, 1268–1275. Conference), pp. 78–79.
A 104, 1876–1889.
Mukherji, S., and Van Oudenaarden, A. (2009). Shen, Y., Chen, Y., Wu, J., Shaner, N.C., and
Greenbaum, L., Rothmann, C., Lavie, R., and Synthetic biology: understanding biological Campbell, R.E. (2017). Engineering of mCherry
Malik, Z. (2000). Green fluorescent protein design from synthetic circuits. Nat. Rev. Genet. variants with long Stokes shift, red-shifted
photobleaching: a model for protein damage by 10, 859–871. fluorescence, and low cytotoxicity. PLoS One 12,
endogenous and exogenous singlet oxygen. Biol. 1–14.
Chem. 381, 1251–1258. Nemazanyy, I., Montagnac, G., Russell, R.C.,
Morzyglod, L., Burnol, A.F., Guan, K.L., Pende, Stephens, K., Pozo, M., Tsao, C.Y., Hauk, P., and
Kim, J., Woo, S.S., and Sarpeshkar, R. (2018). Fast M., and Panasyuk, G. (2015). Class III PI3K Bentley, W.E. (2019). Bacterial co-culture with cell
and precise emulation of stochastic biochemical regulates organismal glucose homeostasis by signaling translator and growth controller
reaction networks with amplified thermal noise in providing negative feedback on hepatic insulin modules for autonomously regulated culture
silicon chips. IEEE Trans. Biomed. Circuits Syst. 12, signalling. Nat. Commun. 6, 8283. composition. Nat. Commun. 10, 1–11.
379–389.
Nielsen, A.A.K., Der, B.S., Shin, J., Stricker, J., Cookson, S., Bennett, M.R., Mather,
Klavins, E. (2014). Lightening the load in synthetic Vaidyanathan, P., Paralanov, V., Strychalski, W.H., Tsimring, L.S., and Hasty, J. (2008). A fast,
biology. Nat. Biotechnol. 32, 1198–1200. E.A., Ross, D., Densmore, D., and Voigt, C.A. robust and tunable synthetic gene oscillator.
(2016). Genetic circuit design automation. Nature 456, 516–519.
Lee, J.W., Gyorgy, A., Cameron, D.E., Pyenson, Science 80, 352.
N., Choi, K.R., Way, J.C., Silver, P.A., Del Vecchio, Sturm, O.E., Orton, R., Grindlay, J., Birtwistle, M.,
D., and Collins, J.J. (2016). Creating single-copy Okamoto, A., Hashimoto, K., Nealson, K.H., and Vyshemirsky, V., Gilbert, D., Calder, M., Pitt, A.,
genetic circuits. Mol. Cell 63, 329–336. Nakamura, R. (2013). Rate enhancement of Kholodenko, B., and Kolch, W. (2010). The
bacterial extracellular electron transport involves mammalian MAPK/ERK pathway exhibits
Li, H., and Petzold, L. (2010). Efficient bound flavin semiquinones. Proc. Natl. Acad. Sci. properties of a negative feedback amplifier. Sci.
parallelization of the stochastic simulation U S A 110, 7856–7861. Signal. 3, 1–8.
algorithm for chemically reacting systems on the
graphics processing unit. Int. J. High Perform. Teo, J.J.Y. (2019). Synthetic Analog Feedback
Okamoto, A., Kalathil, S., Deng, X., Hashimoto, K.,
Comput. Appl. 24, 107–116. Control Circuits in Living Cells, Ph.D Thesis
Nakamura, R., and Nealson, K.H. (2014). Cell-
secreted flavins bound to membrane cytochromes (Massachusetts Inst. Technol).
Lichtman, J.W., and Conchello, J.A. (2005).
dictate electron transfer reactions to surfaces with
Fluorescence microscopy. Nat. Methods 2,
diverse charge and pH. Sci. Rep. 4, 1–8. Teo, J.J.Y., Woo, S.S., and Sarpeshkar, R. (2015).
910–919.
Synthetic biology: a unifying view and review
Opgenorth, P., Costello, Z., Okada, T., Goyal, G., using analog circuits. IEEE Trans. Biomed.
Lisa, G., Paul, F., and Polizzi, K.M. (2014). Co-
Chen, Y., Gin, J., Benites, V., De Raad, M., Circuits Syst. 9, 453–474.
culture systems and technologies: taking
Northen, T.R., Deng, K., et al. (2019). Lessons from
synthetic biology to the next level. J. R. Soc.
two design-build-test-learn cycles of dodecanol Teo, J. Kim, J. Woo, S. and Sarpeshkar, R. ‘‘Bio-
Interfaces 11, 20140065.
production in Escherichia coli aided by machine molecular Circuit Design with Electronic Circuit
learning. ACS Synth. Biol. 8, 1337–1351. Software and Cytomorphic Chips’’, Proceedings of
Liu, Z., Gerner, M.Y., Van Panhuys, N., Levine, A.G.,
Rudensky, A.Y., and Germain, R.N. (2015). Immune the IEEE Biomedical Circuits and Systems
homeostasis enforced by co-localized effector and Pan, D., Fujimoto, M., Lopes, A., and Wang, Y.X. Conference (BioCAS 2019), Nara, Japan, pp. 1–4
regulatory T cells. Nature 528, 225–230. (2009). Twist-1 is a PPARd-inducible, negative-
feedback regulator of PGC-1a in Brown fat Wong, V.W.Y., Stange, D.E., Page, M.E.,
Lyons, A.B., and Parish, C.R. (1994). metabolism. Cell 137, 73–86. Buczacki, S., Wabik, A., Itami, S., Van De
Determination of lymphocyte division by flow Wetering, M., Poulsom, R., Wright, N.A., Trotter,
cytometry. J. Immunol. Methods 171, 131–137. Prindle, A., Selimkhanov, J., Li, H., Razinkov, I., M.W.B., et al. (2012). Lrig1 controls intestinal
Tsimring, L.S., and Hasty, J. (2014). Rapid and stem-cell homeostasis by negative regulation of
Zeng, J., Teo, J.J.Y., Banerjee, A., Chapman, tunable post-translational coupling of genetic ErbB signalling. Nat. Cell Biol. 14, 401–408.
T.W., Kim, J., and Sarpeshkar, R. (2018). A circuits. Nature 508, 387–391.
synthetic microbial operational amplifier. ACS Woo, S.S. (2016). Fast Simulation of Stochastic
Ramaswamy, R., González-Segredo, N., and Biochemical Reaction Networks on Cytomorphic
Synth. Biol. 7, 2007–2013.
Sbalzarini, I.F. (2009). A new class of highly efficient Chips, PhD. Thesis (Massachusetts Inst. Technol.).
Mandal, S., and Sarpeshkar, R. (2009a). Log- exact stochastic simulation algorithms for chemical
domain circuit models of chemical reaction networks. J. Chem. Phys. 130, 1–14. Woo, S.S., Kim, J., and Sarpeshkar, R. (2015). A
reactions. In 2009 IEEE International cytomorphic chip for quantitative modeling of
Symposium on Circuits and Systems, (IEEE), Sarpeshkar, R. (2010). Ultra Low Power fundamental bio-molecular circuits. IEEE Trans.
pp. 2697–2700. Bioelectronics: Fundamentals, Biomedical Biomed. Circuits Syst. 9, 527–542.
Applications, and Bio-Inspired Systems
Mandal, S., and Sarpeshkar, R. (2009b). Circuit (Cambridge University Press). Woo, S.S., Kim, J., and Sarpeshkar, R. (2018). A
models of stochastic genetic networks. In 2009 digitally programmable cytomorphic chip for
IEEE Biomedical Circuits and Systems Sarpeshkar, R. (2014). Analog synthetic biology. simulation of arbitrary biochemical reaction
Conference, (IEEE), pp. 109–112. Philos. Trans. R. Soc. A. Math. Phys. Eng. Sci. 372, networks. IEEE Trans. Biomed. Circuits Syst. 12,
20130110. 360–378.
Maung, N.W., and Smolke, C.D. (2008). Higher-
order cellular information processing with Sarpeshkar, R., Salthouse, C., Sit, J.J., Baker, Zeng, J., Banerjee, A., Kim, J., Deng, Y.,
synthetic RNA devices. Science 322, 456–460. M.W., Zhak, S.M., Lu, T.K.T., Turicchia, L., and Chapman, T.W., Daniel, R., and Sarpeshkar, R.
Balster, S. (2005a). An ultra-low-power (2019). A novel bioelectronic reporter system in
Medley, J.K., Teo, J., Woo, S.S., Hellerstein, J., programmable analog bionic ear processor. IEEE living cells tested with a synthetic biological
Sarpeshkar, R., and Sauro, H.M. (2020). A Trans. Biomed. Eng. 52, 711–727. comparator. Sci. Rep. 9, 1–7.
Supplemental Information
We mapped the Bio-OpAmp to its analog circuit equivalent using the mapping discussed in the
previous section, and as shown in Fig. 4c. Currents represent molecular fluxes while node
voltages represent molecular concentrations. We model transcription and translation using
dependent current generators connected in parallel with a resistor and capacitor (RC). The
current generators vary their output depending on their voltage inputs, which mimics
transcription and translation regulation. This trio of analog parts shown describes a simplified
transcription and translation process which we use to model the dynamics of sgRNA, dCas9,
LuxI, AiiA and AHL.
Low sensitivity to parameter variation can similarly be achieved with a Bio-OpAmp. In this case,
open-loop gain is a function of catalytic efficiency, dissociation constant, degradation rate and
the steady-state concentration of the underlying biomolecules. By applying small-signal circuit
analysis to the analog circuit in Supplementary Fig. 4 (Sarpeshkar, 2010), we recover the precise
relationship between loop gain and biological parameters (See (Zeng et al., 2018)). Since an
OpAmp is robust to variations in open-loop gain if its value is much greater than 1, we can ‘spread’
the uncertainty to parameters that are difficult to control and tune the controllable parameters
to generate a large overall loop gain. One other important consideration that applies specifically
to the Bio-OpAmp is that biochemical reactions are generally sigmoidal. If the concentration of a
particular biomolecule is operating close to saturation, it responds more slowly to change,
therefore reducing loop gain and attenuating negative feedback. To reduce sensitivity, the Bio-
OpAmp should operate at a level where its constitutive molecules have concentrations that are
close to their respective dissociation constants in order to maximize open-loop gain.
A classical OpAmp is often employed in electronic circuits because of its resistance to upstream
and downstream loading effects i.e. connecting an OpAmp to an existing circuit does not affect
either the circuit or the OpAmps original behavior. In synthetic biology, loading effects are
often called retroactivity and they are known for introducing context-dependent effects to
otherwise well characterized synthetic circuits (Del Vecchio et al., 2008). For example, a
repressor in a naïve feedback circuit binds to an upstream promoter to create negative
feedback (Stricker et al., 2008), but a cognate promoter downstream that is placed on a high
copy plasmid acts like a shunt that competes for the same repressor, thereby attenuating the
strength of the negative feedback. In contrast, the Bio-OpAmp rejects the downstream
disturbance by leveraging its high open-loop gain to attenuate the ‘error’ such that the output
is determined only by input Arabinose. Additionally, the third gain stage is designed to be fast
so that LuxI and AiiA can respond to downstream changes of AHL quickly. Fast dynamics,
combined with high open-loop gain, are responsible for the low output impedance of the Bio-
OpAmp. As for the input signal, the low copy number of araC in the first stage of the Bio-
OpAmp ensures that the upstream concentration of Arabinose is not affected by its use-it-and-
lose-it binding. Low araC copy number is therefore responsible for the high input impedance of
the Bio-OpAmp
One potential problem with multi-stage OpAmps in negative- feedback loops is the occurrence
unintended oscillation caused by large open-loop gains. In Bio-OpAmps, the problem is further
compounded by delays from binding, transcription and translation initiation and diffusion of
molecules. Intuitively, delays and phase shifts are well-known for causing oscillations in
otherwise stable systems because they prime a feedback loop to overcorrect an initial
disturbance due to a delay in response, and the overcorrection triggers another round of
correction in the opposite direction through feedback, therefore creating oscillation
(Sarpeshkar, 2010). To attenuate oscillations, we could reduce open-loop gain at the expense of
increasing tracking error and reducing OpAmp performance. An alternative solution commonly
employed in analog electronic circuit design is to introduce a dominant pole by slowing the
time scale of one of the stages. We achieve the slowdown in the second stage of the Bio-
OpAmp by leveraging the relatively slow process of protein translation compared to RNA
transcription and small-molecule synthesis.
When designing the Bio-OpAmp, we modeled and simulated the analog circuit equivalent (see
Fig. 4) with the circuit software tool, Cadence, a computer aided design (CAD) tool. Cadence
converts the schematic into a system of discrete-difference numerical equations that it
simulates on a CPU through well-known and fairly advanced numerical-analysis techniques,
honed over decades to improve convergence. It derives implicit relationships between voltages
and currents by applying the laws of conservation of charge so users can avoid the error-prone
process of accounting for reaction fluxes manually. We can change the topology simply by
connecting wires between circuit motifs and replacing components with the click of a button.
Using the suite of analysis tools in Cadence, we performed frequency analysis to identify
sources of instability in Bio-OpAmp and eliminated them using the dominant-pole
compensation technique from analog circuit design (Sarpeshkar, 2010). We were able to arrive
at our final Bio-OpAmp design because we adopted an analog circuit framework and were able
to leverage its many tools and concepts.
Supplementary Section II: Compiling Circuits
To simulate any transcription-translation network on the transistor-based cytomorphic chips,
we first create a mapping for the RC circuit motif of the Bio-OpAmp (See Supplementary Fig. 6
and Supplementary Fig. 7) (Sarpeshkar, 2010, 2014; Teo et al., 2015). Each cytomorphic block
on the chips is designed to produce a current flux given by a general bidirectional (A + B) ->
(C+D) reaction with C being capable of being independently produced or degraded by other
reactions. It is parametrized and modeled by the following equation:
𝑛
𝐵𝑡𝑜𝑡 − 𝐶𝑡𝑜𝑡 𝐶𝑓𝑟𝑒𝑒 𝐷𝑓𝑟𝑒𝑒
Rate = 𝑘𝑟 (((𝐴𝑡𝑜𝑡 − 𝐶𝑡𝑜𝑡 ) ( ) + 𝐶𝑝𝑟𝑜𝑑 ) − ( ))
𝐾𝐷𝑓𝑤 𝐾𝐷𝑟𝑣
where 𝐴𝑡𝑜𝑡 − 𝐶𝑡𝑜𝑡 = 𝐴𝑓𝑟𝑒𝑒 and 𝐵𝑡𝑜𝑡 − 𝐶𝑡𝑜𝑡 = 𝐵𝑓𝑟𝑒𝑒 if C (or D) can only be produced by a
reaction involving only A and B (Woo et al., 2018). Supplementary Fig. 7 shows how the
cytomorphic circuit building blocks can be wired to simulate a transcription-translation
reaction. The configuration shown here can be viewed as a variant of the Michaelis-Menten
(MM) enzyme-substrate binding configuration for protein blocks that is described in Fig. 2 of
(Woo et al., 2018): It assumes that the substrate concentration is much larger than enzyme
concentration such that substrate depletion is negligible. This assumption works because the
number of transcription factors (the ‘substrate’) is much larger than the number of binding sites
for the transcription factor on DNA (the ‘enzyme’ ). To emulate the lack of substrate depletion
explicitly, we open the use-it-and-lose-it feedback loop that connects Ctot to Btot. To reduce the
number of parameters and improve model interpretability, we typically use normalized
variables, i.e., variables wherein substrate concentrations are divided by their respective
dissociation constants.
𝑋
K dX
IX = Etot
S1 S2 Eq. 2
1+K +K
dS1 dS2
where 𝑋 is either S1 and S2. To model the dynamics of enzyme substrate binding, we can add a
capacitor in parallel to the conductance divider. To model competition between more substrate
species, we simply add a conductance for each additional substrate. The circuit can be mapped
onto the cytomorphic chip using a variant of the fan-out configuration described in Fig. 7b of
(Woo et al., 2018). Just like the transcription-translation motif, we assume that S >> E so we can
ignore substrate depletion by removing the use-it-and-lose-it feedback loop, as shown in
Supplementary Fig. 8. To increase the number of competing substrates, we simply add a new
protein block for each additional substrate and subtract bound ES complexes from ETot.
Finally, we need a circuit motif to model the third differential gain stage of the Bio-OpAmp of Fig.
4. The enzymes, LuxI and AiiA synthesize and degrade the AHL molecule respectively, which is
modeled by using a current generator that is controlled by LuxI and a degradation resistor that is
controlled by AiiA as shown in Supplementary Fig. 9a. The cytomorphic instantiation as shown in
Supplementary Fig. 9b converts the reverse reaction from Eq. 1 to a degradation reaction.
As we alluded to in the previous section, the second amplification stage utilizes the competition
between sgRNA1 and sgRNA2 for dCas9 to increase gain. To optimize the strength of
amplification, we performed a 2D sweep on the dissociation constants between sgRNAs and
dCas9. Supplementary Fig. 12 shows that differential gain is maximized when both dissociation
constants are the same. When one of the sgRNA concentration increases while the other stays
constant, it not only forms more complexes with dCas9, it also displaces the other sgRNA from
dCas9, thus creating a push-pull reaction. If the binding affinities of both sgRNAs are similar, the
symmetry ensures that there is an equal concentration of dcas9-sgRNA1 and dCas9-sgRNA2 such
that any change in either sgRNA concentrations will elicit a stronger push-pull effect than if the
concentration of complexes were disproportionate. Another observation is that a stronger
binding affinity increases differential gain. In the presence of a large concentration of unbound
dCas9, increases in either sgRNA concentrations would increase binding to unbound dCas9 copies
instead of the bound complexes. The unbound dCas9 therefore acts as a buffer that attenuates
gain. By increasing binding affinity, we ensure that most dCas9 are bound to sgRNA, therefore
reducing their buffering capacity.
Supplementary Fig. 1. Dynamics of the circuit of Fig. 2 that models an enzyme-substrate
binding reaction. Related to Section I. The circuit reproduces the dynamics of an enzyme-
substrate binding reaction. Voltages are analogous to molecular concentration and represented
in green text. Currents are analogous to molecular flux. Using Kirchhoff’s current law, the
current flowing into the capacitor, Ic, is the difference between Iasc and Idis. If the rate of
association is greater than the rate of dissociation, a positive Ic flows into the capacitor (C) and
causes an accumulation of charge (q) which increases the voltage ES. When Ic is zero, ES
stays constant and the system is in equilibrium. Iasc is a second-order association reaction
which depends on the product of voltages Efree and Sfree. We use voltage adders to generate the
free quantities and a voltage multiplier to calculate their product. The dependent current
generator, represented by a diamond symbol, scales the input voltage by Kf to produce Iasc. Idis
is a first-order reaction representing dissociation. We model the first-order behavior using a
simple resistor that depends on the reciprocal of Kr.
Supplementary Fig. 2. Canonical Analog Circuit that represents DNA, RNA, Protein,
and Small-Molecule Circuits in Molecular Biology. Related to Section I. The canonical
circuit schematic represents biological reactions involving DNA (blue), RNA (red), protein
(green) and small-molecules/metabolites (purple). It contains sub-circuits for transcription,
translation and various forms of regulations including ribosomal switches, transcription
factors, splicing, post-translational modification and CRISPR-based regulation. Portions of
this circuit have been instantiated in 17 synthetic biological circuits over 20+ years and have
shed insight into their operation (Teo et al., 2015).
Supplementary Fig. 3. Cadence – An analog electronic circuit Computer Aided Design
(CAD) software tool. Related to Section II. This software tool includes a suite of 26+ forms
of analysis that have been used over decades to design extremely complex integrated circuits
with billions of transistors that meet important system performance specifications involving
speed, accuracy, power efficiency, robustness, and device count or chip size.
a)
b) c)
b)
c)
Supplementary Fig. 6. Examples of multiple configurations of protein blocks on
cytomorphic chips that can simulate various biochemical reactions. Related to Section
III and Fig. 6. a) Each chip contains multiple protein blocks with current input and output
ports as shown in the figure. Embedded within the blocks are multiple switches and Digital-
to-Analog Converters (DACs) to customize biochemical network parameters. Back-
propagation of downstream effects to upstream reactions due to use-and-lose-it feedback
loops in non-modular cascades of reactions are also accurately represented. b) To simulate a
fan-out reaction, we connect two protein blocks as shown in (Woo et al., 2018). The first
block simulates the binding between A and B. The second block accounts for binding
between A and C. c) To simulate an enzyme-substrate catalytic reaction, one block is used to
model the formation of the enzyme-substrate complex. A second block captures the molecular
dynamics and molecular production/conservation inherent in substrate-to-product conversion.
a)
b)
b)
Supplementary Fig. 8. Simulation of a fanout circuit motif block with cytomorphic-chip
protein blocks. Related to Section III and Fig. 6. (a) The conductance-divider circuit motif
shown above is the same as that in the microbial Op-Amp circuit model corresponding to Fig.
4c that models competitive binding of dCas9 to sgRNA1 or sgRNA2. Each conductance
represents the normalized concentration of sgRNA1 or sgRNA2, and the current flux through
each conductance represents the amount of bound dCas9 respectively. A third conductance of
value 1 accounts for the proportion of unbound dCas9. (b) Each cytomorphic circuit block
calculates the amount of enzyme bound to either sgRNA1 (dCs1) or sgRNA2 (dCs2)
respectively, accounting for accumulating depletion in dCas9 when bound, in a cascade.
fashion.
a)
b)
Supplementary Fig. 13. Closed-loop gain with varying gain of Arabinose and AHL.
Related to Section III. a) Circuit software simulation. b) Cytomorphic chip simulation. In the
Bio-OpAmp, the feedback loop gain is determined the small-signal gain of AHL/Arabinose.
The Closed-loop gain of the Bio-OpAmp changes in proportion to the reciprocal of the
feedback loop gain.
Supplementary Fig. 14. Steady-State Error vs Gain in a Bio-OpAmp. Related to Section
III. The steady-state tracking error of a closed-loop Bio-OpAmp scales like 1/(1 + 𝐺) in the
presence of negative feedback, where 𝐺 is the differential gain of the operational amplifier. As
gain increases, the steady-state error drops as predicted and confirmed via simulations in
software and on cytomorphic chips.